The same edit, three routes, wildly different clocks. The fast paths are genuinely usable, you’re trading faithfulness to the original composition, not raw quality.
| Route | Measured | Character |
|---|---|---|
| Qwen + 4-step Lightning LoRA | 15s | Conservative. Small surgical edits. |
| Flux.1 Kontext | 18s | Fast but looser, it changed our jacket’s cut and nudged the face. |
| Qwen-Image-Edit-2511, full | 57s | Faithful. Big changes without drift. |
Follow-up, September 2026: we re-ran this on five current editors, took two people out of a group photo, and took back one thing this guide’s era said was impossible. Read Remove People From a Photo With Free Local AI and the companion piece on why editors keep faces and lose bodies.
Wiring the fast path
One node inserted, two numbers changed.
LoraLoaderModelOnly goes between your UNETLoader and the KSampler, with the 4-step Lightning LoRA
selected at strength 1.0. Then:
| Field | Normal | Lightning |
|---|---|---|
steps | 20 | 4 |
cfg | 2.5 | 1.0 |
Leaving CFG at 2.5 with a 4-step LoRA is the usual mistake, it produces a burnt, over-contrast image. The accelerator expects CFG at or near 1.
Which to reach for
- Iterating on wording? Lightning. Fifteen seconds means you can try eight phrasings in two minutes.
- Committed to the shot? Full mode. Take the fifty-seven seconds.
- Need a different model’s opinion? Kontext, but check the face afterwards.
Two more things worth knowing
Changing the camera angle is real and cheap. Asking for a three-quarter side view returned the same person, same room, same clothing, from a genuinely different angle in 42 seconds.
Text editing has a hard limit. It writes requested words correctly on flat, front-facing surfaces. On a tilted sheet of paper it rendered them mirrored. Front-facing only.
The settings, and what each one actually does
Four numbers in the KSampler carry an edit. These are the values we ran for every result below.
| Field | Value | Why |
|---|---|---|
steps | 20 | Below about 12 it stops respecting the instruction properly. |
cfg | 2.5 | The one people get wrong. On an edit model, high CFG doesn’t mean “obey me more”, it means “destroy the original.” Stay between 2 and 3. |
sampler_name | euler | Boring and reliable. |
scheduler | simple | Same. |
denoise | 1.0 | Surprising, but correct here: the encoder holds your original, not the noise level. |
seed | change it | When an edit half-works, same prompt + new seed is free. |